What is generative engine optimisation (GEO)?

Generative engine optimisation is the practice of making your business visible and recommendable to AI systems — ChatGPT, Claude, Perplexity, Gemini — when people ask them for help in your category. It is the equivalent of what SEO did for search engines, but the mechanics are different enough that treating it as "SEO with a new name" will cost you.

The difference is not cosmetic. It changes what winning looks like.

How is GEO different from SEO?

Search engines return a list. Ten links, sometimes more, and being fourth is still a business. Someone scanning results will see you, and a share of them will click. Position matters, but the list is generous enough that presence is worth something on its own.

Generative systems do not return a list. They return an answer, and that answer names one option, sometimes three, occasionally none at all. There is no page two. There is no scrolling to find you further down. You are either in the answer or you are not in the conversation.

That has a second consequence people underestimate: the user may never visit your site. If someone asks an assistant which firms handle AI governance for broadcasters and it names three, the value you received was being named — not the click. Traditional analytics will not show it. You may be winning or losing at this right now with no signal either way in your dashboard.

So the objective shifts. SEO asks: how do we rank for this query? GEO asks: when a model is asked this question, does it know we exist, does it understand what we do, and does it trust the information enough to say our name?

Why does this matter now?

Because the behaviour has already moved. People increasingly open an assistant rather than a search box for the exploratory, high-consideration questions — "who should I talk to about X", "what are the options for Y" — which are precisely the questions that precede a purchase in B2B.

This is also the point where the honest caveat belongs: nobody has a definitive map of how generative systems weight sources. There is no published algorithm, no equivalent of decades of SEO experimentation, and any vendor claiming precise ranking factors is selling you certainty they do not have. What we can describe with confidence is the mechanism — how these systems acquire information about a business, and what makes that information usable. That is enough to act on, and it is where the rest of this piece stays.

What actually drives whether an AI recommends you?

Models learn about your business two ways: from training data, which is historical and slow and largely outside your control, and from retrieval at query time — searching, browsing, reading pages. The second is where your work pays off.

Say plainly what you do. This sounds too simple to matter. It is not. A model categorising your business reads the top of your page looking for a statement it can bind to your name: this organisation does this specific thing for these specific people. If your homepage opens with a concept, a value proposition or a clever line, a human who already knows your category will follow it — and a model will come away with nothing it can confidently repeat.

I found this on my own site. Zapheron's homepage led with a headline I still like, and underneath it, language that described a philosophy rather than a service. Our own scoring tool flagged it. The fix was one clause: the paragraph now begins by naming the company, the audience and the work. That is not a copywriting flourish; it is giving a machine something factual to hold onto.

Make the information structured. Schema markup — Organization, Person, FAQPage — is not a ranking trick. It is the difference between a model inferring who you are from prose and being told directly, in a format built for the purpose. Entity clarity compounds: an Organization block that names a founder, linked to a Person block with credentials and verifiable external profiles, gives a model a coherent picture rather than scattered mentions.

Be readable by machines. A surprising number of sites fail here for boring reasons: no sitemap, no robots.txt, content that only appears after JavaScript runs. Many crawlers do not execute JavaScript. If your content assembles itself in the browser, it may simply not exist to the systems you are trying to reach.

Get described by other people. This is the uncomfortable one, because you cannot ship it on a Friday. Models weight third-party corroboration heavily — reviews, case studies, named client quotes, coverage, mentions in places you do not control. What you say about yourself is a claim. What others say about you is evidence. Any credible GEO effort eventually runs into the fact that you have to earn some of this rather than build it.

Answer questions in the shape they are asked. Content organised around real questions — as headings, answered directly beneath — matches the structure of what users type into an assistant. This is also why FAQ markup earns its place: it hands the model a question and an answer already paired.

Where should you start?

In order, because the sequence matters:

1. Fix the plain statement of what you do, in the first visible paragraph of your homepage. Cheapest change on this list, and frequently the one holding everything else back.
2. Add structured data — Organization at minimum, then Person for a founder-led business, then FAQPage.
3. Check you are machine-readable — sitemap, robots.txt, and content in the HTML rather than assembled client-side.
4. Build question-shaped content on the topics you want to be recommended for.
5. Then work on third-party validation, which is slower and mostly not a website task at all.

Most organisations discover that steps one to three are a week of work that has simply never been prioritised, and that they were invisible for reasons entirely within their control.

If you want a concrete starting point, our AI Visibility Scorer runs 21 checks against your site and returns a score with specific fixes. It is free, it takes seconds, and it will tell you which of the above you are already failing.

Fair warning: we ran it on our own site and scored 83.